What happened
Taiwan’s Central News Agency (CNA) has launched a subscription service built for AI, called CNA MCP. Once you subscribe, any news-related question you put to an AI assistant such as ChatGPT or Claude gets looked up directly in CNA’s news archive in real time, and the answer is required to cite its source. It opened to subscribers on 31 August 2026 at NT$200 a month.
The MCP in the name is a standard way of connecting an AI assistant directly to a database so it can look things up — think of it as a dedicated line running from the AI to the archive. The line doesn’t appear on its own. Subscribers have to attach it to the assistant they already use.
The archive holds nearly 5 million news stories going back more than 30 years, along with news photos, world yearbook data and public data from government agencies. CNA says it is the first news organisation in Taiwan to let AI look up its reporting through MCP, and that anyone from government bodies and schools to individuals can subscribe.
Why do it? CNA’s president framed it around “zero-click”: more and more people ask AI directly and never visit a website. The executive who led the project put it more starkly in an interview with the Taiwanese tech magazine Business Next: “Because the end of websites is coming.” He also admitted it’s too early to say whether MCP will become the next main gateway.
Why it matters to you
What CNA has changed is how AI gets its material. Until now, when an AI answered a question about Taiwan, it gathered that material itself from the open web. It sends out programs that fetch web pages automatically — these are called crawlers — and whichever version a crawler happened to pick up, from whichever year, is the version the answer reflects. CNA has stopped waiting. Once a user subscribes, it sends the version it reported and edited itself straight into the conversation, with the source printed on the answer.
So AI now has two ways of getting information. One is what CNA has done: plug its material into the AI directly, so the AI finds it at once and cites it, with users paying for the privilege. The other is the ordinary public web page, which has to wait for a crawler to pass by and pick it up. Most Taiwanese brands’ websites and official information are still in that second group.
If you’ve never been in a CNA story, the first route doesn’t touch you. The second one does. When customers ask an AI about your business, it still answers, using whatever it picked up on the web, and that often isn’t the version you wrote. You can check this yourself. Open ChatGPT, ask what your company does and what’s new with it, and look at what it cites. Often it’s a news piece, a comparison site or a forum thread, sometimes an article several years old — rarely the version on your own site.
Say you run a bakery that moved three years ago. The food blogger’s write-up from back then still lists the old address, and the price-comparison site still shows the signature products you sold the year before last. Whichever page the AI picks up is the one it repeats to customers; the new address on your own site may never have crossed its path.
It works the other way too. If you have been in a CNA story, and someone with CNA MCP attached asks about you, the AI may pull that story straight up and put it in the answer with a citation, whether it covered a product launch or a customer complaint.
AI doesn’t only describe you from other people’s versions; sometimes it gets the names wrong too. The project lead noted that AI sometimes picks up the translations of names used in mainland China; with CNA MCP attached, answers stay close to the terms and names people in Taiwan actually use. So for subscribers, at least, names in the news now have CNA checking them. Nobody is checking how AI writes your brand name, your product names or your company’s history.
At NT$200 a month, and with users having to connect it themselves, this won’t be on everyone’s phone any time soon. Going by the price and the setup, most subscribers will probably be people who look things up all day — researchers, schools, public agencies — so it reaches few people for now. What’s worth watching is the direction it points in: sources with real weight have started feeding their content into AI themselves, while everyone else’s information still waits for AI to find it on the web.
What to do about it
Act now, though not by building an MCP of your own. What lets CNA charge is that archive of 30 years of verified reporting, the kind of volume people will pay to plug into their AI. Nobody is going to subscribe to a company website.
Your concern is the second route: when the AI goes out and picks things up on its own, what it picks up should be your official version. For example, your founding year, shop address and signature products should read the same on your site, in press coverage and on other people’s pages about you. Your brand and product names should be written one consistent way in Chinese, not one way here and another way there. And your site’s content has to be something AI’s crawlers can reach and actually make sense of.
The trouble is that the pages AI cites when it talks about you are mostly not yours. The news piece, the comparison site, the forum thread you saw when you checked earlier all belong to someone else, and you can’t change a word of them. What you can do is find out which pages the AI is citing right now, what each one says about you and which line contradicts your own account, then use your official version to fill in what’s missing and outweigh what’s out of date, so that next time the AI picks up your version. Until you have that list of sources, you can keep editing your site and still be patching places the AI isn’t even reading.
We’re geoweb.tw, and our managed GEO service deals specifically with how AI describes you: our team fixes your site’s content, brings what’s said about you elsewhere on the web into line, and regularly checks how each AI describes you — you just review the results. If you’d like to see which pages AI cites when it talks about you today, and what each one says, book a demo: set up a session with one of our consultants and we’ll ask a few questions live.
Further reading
- Do you have a say in how AI uses your content? — the licensing layer, mapped
- AI has simply never read you: the Traditional-Chinese corpus gap — why AI keeps getting Taiwanese brands wrong
- MCP’s biggest rewrite yet: AI is growing its own forms — where this interface goes next